Computer-aided Cataract Grading Under Adversarial Environment

Cataract is the most common cause of blindness in the world. Early detection and treatment can lower the risk of cataract progression. The diagnostic performance of existing computer-aided cataract grading (CACG) methods often deteriorates due to the sophisticated image capture technology. The common retinal fundus image aberrations such as noise and blur are unavoidable in practice. In this paper, a CACG method is proposed to achieve robust cataract grading under adversarial conditions such as noise and blur. The presented CACG method is designed using three deep neural network variants. Each variant is fine-tuned individually using good, noisy, and blur retinal fundus images to achieve optimum performance. Further, the input image quality detection module is incorporated in the proposed CACG method to detect input image distortion and then pivots the input image to the desired deep neural network variant. Gaussian noise and blur models are used to evaluate the effectiveness of the suggested CACG method. The proposed CACG approach exhibits superior performance to existing methods under adversarial conditions.